Event-independent temporal positioning: application to French clinical text

Nesrine Bannour, Bastien Rance, Xavier Tannier, Aurelie NeveolACL 2023

TL;DR

This paper introduces a novel "event-independent" approach to temporal relation extraction in clinical texts. By classifying text segments based on their temporal relationship to the document's creation time, rather than focusing on specific events, the method simplifies annotation and improves robustness across different clinical domains. This approach achieves high accuracy in French clinical reports and shows promise for building patient timelines.

What Problem Does This Paper Address?

Extracting temporal information from clinical texts is crucial for understanding patient histories and treatment pathways. However, existing methods often struggle with the complexity and domain-specificity of clinical narratives. Temporal expressions can be varied, and events may not always have explicitly stated times. Furthermore, annotating clinical texts is time-consuming and requires medical expertise, leading to poor inter-annotator agreement and limited availability of annotated corpora, especially for non-English languages like French. Traditional approaches often tie temporal relation extraction to specific event types, making them less adaptable to new tasks or domains.

Methods

The core contribution of this paper is an event-independent temporal positioning representation. Instead of identifying specific clinical events and then determining their temporal relationship to the Document Creation Time (DCT), this method segments the text into portions that are temporally homogeneous. Each portion is then classified into one of five categories: TemporalReference, Before, Before_Overlap, Overlap, and After, indicating its temporal relationship to the DCT. This approach simplifies the annotation process, making it faster and more reproducible.

The temporal relation classification task is framed as a sequence labeling task. A neural token classification model, specifically a French CamemBERT model from the HuggingFace transformers library, is trained using the BIO (Beginning-Inside-Outside) tagging scheme. This allows the model to identify tokens that signal a temporal shift within the clinical text.

Datasets

The authors utilize two clinical corpora for their work:

  1. Temporal extraction corpus: This corpus consists of de-identified hospital, operative, and consultation reports from colon cancer patients. It was used for training and validating the temporal positioning model.
  2. Toxicity corpus: This corpus, also from de-identified hospital clinical reports, focuses on toxicity information related to chemotherapy for colon and lung cancer patients. It was used to evaluate the efficacy of the temporal positioning approach on a different clinical task (toxicity event extraction).

The temporal annotation scheme used is based on the THYME-TimeML annotation scheme, with the addition of two categories: TemporalReference and End_Scope.

Key Insights and Ideas

  • Simplifying Temporal Annotation: By shifting focus from specific events to broader text segments, the annotation process becomes significantly less demanding and more consistent, leading to higher inter-annotator agreement.
  • Domain and Task Independence: The event-independent representation is designed to be adaptable across different clinical domains and extraction tasks, overcoming a major limitation of previous methods.
  • Bridging the Gap for Non-English Languages: The work specifically addresses the need for robust temporal information extraction in French clinical texts, a less-explored area.
  • Foundation for Patient Timelines: This approach lays the groundwork for constructing comprehensive patient timelines by providing a structured temporal ordering of medical information.

Key Results

  • The proposed event-independent temporal positioning model achieved an overall macro F-measure of 0.86 on the temporal extraction test corpus, outperforming the baseline model (F-measure of 0.35) and exceeding inter-annotator agreement levels.
  • The model demonstrated good performance across various categories, with high F-measures for prevalent categories like Overlap (0.9) and Before (0.86), and also for less represented categories like TemporalReference (0.91) and After (0.88).
  • When applied to the toxicity corpus, the model achieved an F-measure of 0.8 for temporal positioning of narrative portions.
  • The temporal positioning of chemotherapy toxicity events on the toxicity corpus yielded an F-measure of 0.62.
  • The authors also report a low carbon footprint for training and testing their model (199 g CO2 equivalent).

Limitations and Future Work

  • Corpus Size: The manually annotated corpora used were relatively small, which is acknowledged as a limitation due to the time-consuming nature of temporal annotation.
  • Generalizability to Other Cancers: While the model showed adaptability to lung cancer reports (in the toxicity corpus) despite being trained on colon cancer data, further validation on a wider range of cancer types and clinical report structures is needed.
  • Event Extraction Performance: The direct extraction of toxicity events using a dictionary-based approach (QuickUMLS) yielded a lower F-measure (0.59), indicating that while temporal positioning is strong, event identification itself can be challenging.
  • Future Work: The authors suggest extending this work to other hospital data warehouses, evaluating it on different event definitions, and further developing patient timeline construction. They also highlight the potential of their model as a useful tool for clinicians in analyzing large patient records.

Summary updated

BibTeX
@inproceedings{bannour2023event,
  author = {Nesrine Bannour and Bastien Rance and Xavier Tannier and Aurelie Neveol},
  title = {{Event-independent temporal positioning: application to French clinical text}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure